Abstract:
|
We proposed two generalized propensity score matching methodologies to estimate treatment main and interaction effects under multiple and multivalued treatments scenarios. The first method: non-bipartite greedy matching, was an extension of the greedy matching for two-group. The second method: non-bipartite greedy caliper width matching allowing for a variable number of subjects to be matched in a matched set, was extended from the bipartite greedy caliper width matching. To validate our methods, we compared them with the generalized inverse probability treatment weighting, and stratification on Propensity function methods through a series of Monte Carlo simulations. Our approaches outperformed existing solutions in some situations. Finally, we applied our methods to the National Medical Expenditure Survey data to examine the average main causal effects of smoking and poverty status, as well as their interaction on annual medical expenditure.
|